The present study examines the association of diabetes with BMI (kg/m2) in Asian-Indian and Melanesian Fijian populations sharing a common environment. A population-based survey was used to investigate the risk of diabetes (defined by glycosylated Hb concentration ≥ 6·5 % among participants who denied previous diagnosis of the disease by a medical practitioner) by sex, ethnicity and strata of BMI in a series of age-adjusted logistic regression models. Ethnicity and BMI interactions were compared using WHO and empirically derived BMI cut-off points. Indians had a greater risk (BMI and age adjusted) of undetected diabetes than Melanesians in both males (OR 2·99, 95 % CI 1·73, 5·17; P < 0·001) and females (OR 2·26, 95 % CI 1·56, 3·28; P < 0·001). BMI ≥ 25 to < 30 and ≥ 30 kg/m2 conferred a higher risk of diabetes compared with a BMI ≥ 18·5 to < 25 kg/m2. Risk was higher for males with a BMI ≥ 25 to < 30 kg/m2 (OR 2·35, 95 % CI 1·24, 4·46; P = 0·007) and BMI ≥ 30 kg/m2 (OR 6·08, 95 % CI 3·06, 12·07; P < 0·001) than for females with the same BMI (OR 1·85, 95 % CI 1·11, 3·08; P = 0·027 and OR 2·10, 95 % CI 1·28, 3·44; P = 0·002, respectively). However, the threshold that appeared to differentiate higher risk varied by ethnicity and sex. For Melanesians, BMI thresholds suggested were 25 kg/m2 for males and 32 kg/m2 for females. For Indo-Fijians, these were 24 and 22 kg/m2 for males and females, respectively. Disaggregating by ethnicity and sex, and applying specific evidence-based thresholds, may render BMI a more discriminating tool for assessing the risk of developing diabetes among Fiji adults.